Background technique
With the development of medical imaging technology and deep learning, MRI tumor image is divided using depth learning technology
Class and detection, increasingly by the concern of domestic and foreign scholars (Pereira S, Pinto A, Alves V, et al.Brain
tumor segmentation using convolutional neural networks in MRI images[J].IEEE
transactions on medical imaging,2016,35(5):1240-1251.).Magnetic resonance imaging (MRI) is not necessarily to
Contrast agent is injected, without ionization radiation injury, generates primary three-dimensional cross-sections three-dimensional imaging, high resolution, bone-free artifact interference.But
MRI image data amount is huge, and manual annotation analysis needs the technical ability with profession for guiding, simultaneous with the subjective judgement of doctor,
Time-consuming and repeatability is low.Relative to traditional machine learning method, deep learning is suitble to the processing sample that data volume is big, feature is complicated
This, and model accuracy is high.Volume and network (CNN) model in deep learning, due to its unique network architecture, in image point
It has a high potential in class, image detection, can be used for the extraction of focal area feature in MRI picture.
Now widely used method is sliced to MRI voxel, is handled 2D data, such as 2016 Holland how
Mei Heng university Setio et al. (Setio A A A, Ciompi F, Litjens G, et al.Pulmonary nodule
detection in CT images:false positive reduction using multi-view
convolutional networks[J].IEEE transactions on medical imaging,2016,35(5):
The slice that 9 directions 1160-1169.) are carried out centered on Lung neoplasm carrys out quick obtaining feature, but loses each focal area and cut
The correlation of piece;More features can be retained relative to 2D MRI using 3D MRI voxel, but also bring huge meter simultaneously
Calculation amount (Dou Q, Chen H, Yu L, et al.Automatic detection of cerebral microbleeds from
MR images via 3D convolutional neural networks[J].IEEE transactions on
medical imaging,2016,35(5):1182-1195.).The present maximum and crucial problem of difficulty be how high precision
The feminine gender and Yangxin sample classified to rate in MRI tumor image, while reducing the training time.
Therefore, the strict control training time has very high while accurately extracting MRI tumour figure feature using 3D CNN
Researching value.
Summary of the invention
The purpose of the present invention is carrying out detection classification to 3D MRI tumour voxel, propose to roll up in a kind of deep learning thus and
The MRI tumour voxel detection method of network model framework.
The technical solution adopted by the present invention is that: the MRI tumour voxel detection method based on convolutional neural networks includes following
Step:
Step 1: based on AlexNet model, establishing dual path 3D CNN basic framework of constitutive model, upper and lower path is adopted respectively
With large and small convolution kernel;Step 2: 1 × 1 × 1 convolution kernel is added before second convolutional layer to reduce parameter and calculation amount;Step
Rapid 3: the training on data set using 3D CNN model obtains false positive by comparison of classification and former data is added in false negative sample
Collection;Step 4: changing the full articulamentum of 3D CNN model into convolutional layer, i.e., 3D CNN model is become into 3D FCN model, with this model
Training, obtains final MRI tumour voxel classification model in the data expanded.
Further, the step 1 specifically includes:
First using AlexNet as basic model framework, the part number of plies is deleted, framework main body is connected entirely by 2 convolutional layers, 2
Connect layer, 1 output layer composition.It is sequentially connected ReLU activation primitive, maximum pond layer and LRN after each convolutional layer to standardize layer, 2
A full articulamentum is sequentially connected, and is sequentially connected Dropout layers and objective function layer after the 2nd full articulamentum.Model receives 3D
MRI picture extracts feature using 21 × 21 × 11 big convolution kernel, can sufficiently extract disease as output, the upper pathway of this model
The feature in stove region and surrounding normal tissue;Feature is extracted using 7 × 7 × 3 small convolution kernel in lower path, accurately extracts focal zone
Reduce the redundancy of information while the unique characteristics of domain.
Further, the detailed process of the step 2 are as follows:
A convolutional layer is added before second convolutional layer, convolution kernel size is 1 × 1 × 1, port number utilizes experience
Setting, while the port number of second convolutional layer is reduced, above step will significantly reduce parameter and calculation amount.
Further, the detailed process of the step 3 are as follows:
Step 3.1: using original 3D MRI voxel image as the input through step 1 and step 2 improved model, then carrying out
Random initializtion model parameter when propagated forward, random parameter obey standardized normal distribution.Using Softmax function as target
Function is exported as a result, Softmax function is as follows:
Wherein class y indicates lesion type, i.e., positive or negative, x represents the 3D MRI voxel value of input, θ representative model
In parameter, the advantages of Softmax function category device, is for each class label, can all export a corresponding probability
Value, i.e. P (y=j | x), while the sum of all categories probability value is 1.
Step 3.2: obtaining output result and be compared with authentic signature, it is defeated to return loss function calculating using Softmax
Loss between result and authentic signature out, it is as follows that Softmax returns loss function:
Wherein m indicates that number of samples, k indicate specimen types, and Ι { } indicates that indicator function, J (θ) indicate system loss value.
Step 3.3: backpropagation is carried out, carries out parameter update using stochastic gradient descent method:
I.e. by seeking its partial derivative to loss function, carry out undated parameter for result as gradient value:
Wherein α indicates the step-length that every subparameter updates.
Step 3.4: one threshold values of setting, each Parameters variation amount Δ θ are less than this threshold values, stop undated parameter, otherwise jump
Go to step 3.1.
Step 3.5: model output result being compared with authentic signature, obtains false negative and false positive sample, and original is added
Data set obtains the data set by expansion, while obtaining the 3D CNN model by pre-training.
Further, the detailed process of the step 4 are as follows:
Step 4.1: by most latter two full articulamentum of pre-training 3D CNN model obtained by step 3, change convolutional layer into,
Middle stochastic parameter initialization, and standardized normal distribution is obeyed, 3D CNN model becomes 3D FCN model at this time.
Step 4.2: on the data set expanded by step 3.5, it is trained with the model that this step 4.1 obtains,
Training step refers to step 3, finally obtains MRI tumour voxel classification model.
The beneficial effects of the present invention are:
Traditional artificial analysis annotates MRI image, needs the technical ability with profession for guiding, sentences simultaneous with the subjectivity of doctor
Disconnected, time-consuming and repeatability is low.Computer 2D slice analysis is carried out to MRI image, the correlation between a large amount of lesions can be lost;It is right
MRI image carries out 3D analysis, focus characteristic can be more retained for 2D, but calculation amount is huge.The invention proposes
A kind of MRI tumour voxel detection classification method based on convolutional neural networks, major networks framework are based on double convolution kernel dual paths
The architecture design of fusion using different size of convolution kernel, carries out data fusion using dual path, Optimized model classifying quality,
Sample characteristics can more accurately be extracted and reduce information redundancy.Before extracting feature using 3D convolution kernel, volume 1 × 1 × 1 is added
Product core reduces parameter and dimension, reduces calculation amount and training time.By 3D CNN model, finally full articulamentum becomes convolutional layer, energy
The voxel picture of arbitrary size is enough received as output, accelerates convergence rate.Network training process is based on sample expansion and model
The method of fine tuning is trained using 3D FCN model on the data set of expansion, the generalization ability of model can be improved.It is this
MRI tumor image is detected based on machine vision and depth learning technology, that is, reduces the cumbersome stream of artificial detection analysis
Journey also improves the accuracy rate of detection tumor image, provides technology branch to the automation of future medicine image and intelligent processing
Support.
Specific embodiment
Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention is purged, is complete
Whole description.
Fig. 1 is model framework schematic diagram, the architecture design based on the fusion of double convolution kernel dual paths;When dimensionality reduction and reduction training
Between be based on 1 × 1 × 1 convolution kernel and FCN network;Improve model generalization ability mainly use data sample expansion and model it is micro-
It adjusts;It is final to realize that the MRI tumour voxel based on convolutional neural networks monitors analysis.
Step 1: building the network architecture model of double convolution kernel dual paths, design convolution kernel size, be primarily used to extract
Focus characteristic and focal area are detected.
(1) dual path network model framework is built
Based on AlexNet network model, the part number of plies is deleted, reduces network model depth, establishes dual path 3D
CNN network model basic framework, model by 2 convolutional layers, 2 activation primitive layers, 2 pond layers, 2 standardization layers, 2 it is complete
Articulamentum, 1 Dropout layers and 1 objective function layer are successively constituted.Wherein activation primitive uses ReLU;Pond layer is using most
Great Chiization can extract most significant feature in characteristic pattern;Standardization is standardized using LRN, and parameter setting is as follows, and k=2 is super
Parameter determines that n=5 is the side length in summation section, α=10 by biasing in prototype-4For zoom factor, β=0.75 is exponential term, all
For the hyper parameter of local acknowledgement's standardized operation;The random deactivating layer parameter of Dropout is set as δ=0.5.
(2) dual path convolution kernel size is set
Consider that general MRI voxel size is 512 × 512 × 150, so upper pathway extracts focal area using big convolution kernel
The features such as size, shape, retain more with the contiguity of normal surrounding tissue, convolution kernel size is set as 21 × 21 × 11.
The feature of lesion itself, such as texture, color feature are extracted using small convolution kernel in lower path, it is possible to reduce the redundancy of information, volume
Product core size is set as 7 × 7 × 3.
Step 2: 1 × 1 × 1 convolution kernel is added before second convolutional layer to reduce parameter and calculation amount.
In view of exporting multiple 3D characteristic patterns after first LRN standardization layer, and second 3D convolutional layer is also multichannel,
A large amount of parameter and calculation amount will be generated by carrying out convolution operation.So in first LRN standardization layer and second 3D convolutional layer
A new convolutional layer is added between convolutional layer, convolution kernel size is 1 × 1 × 1, increases the port number of this layer, reduces second
The port number of convolutional layer reduces calculation amount while accurate extraction feature to reach.
Step 3: on raw data set, 3D CNN model improved to step 2 carries out pre-training, passes through comparison of classification
It obtains false positive and original data set is added in false negative sample.
(1) it is based on Xavier parameter initialization method
Network parameter carries out random initializtion, and it is 0 that random parameter, which obeys mean value, and the standard gaussian that variance is 1 is distributed.This can
So that the expectation of network stabilization is consistent after the expectation with training of network parameter when starting to train, for the side for keeping data distribution
It is poor not change with the number of input neuron, using based on Xavier parameter initialization, it is assumed that s is being somebody's turn to do without nonlinear change
Layer network output is as a result, concrete analysis is as follows:
Wherein ω is the layer parameter, and x is this layer of input data.
(2) propagated forward
Using original 3D MRI voxel image as the input through step 1 and step 2 improved model, to biography before then carrying out
It broadcasts.It is exported as objective function as a result, Softmax function is as follows using Softmax function:
Wherein class y indicates lesion type, i.e., positive or negative, x represents the 3D MRI voxel value of input, θ representative model
In parameter, the advantages of Softmax function category device, is for each class label, can all export a corresponding probability
Value, i.e. P (y=j | x), while the sum of all categories probability value is 1.
(3) loss function is calculated
Obtain output result be compared with authentic signature, using Softmax return loss function calculate output result and
Loss between authentic signature, it is as follows that Softmax returns loss function:
Wherein m indicates that number of samples, k indicate specimen types, and Ι { } indicates that indicator function, J (θ) indicate system loss value.
(4) backpropagation
Parameter update is carried out using stochastic gradient descent method:
I.e. by seeking its partial derivative to loss function, carry out undated parameter for result as gradient value:
Wherein α indicates the step-length that every subparameter updates.
(5) repetitive exercise
A threshold values is set, each Parameters variation amount Δ θ is less than this threshold values, stops undated parameter, before otherwise carrying out again
To propagation.
(6) data extending
Model output result is compared with authentic signature, false negative and false positive sample is obtained, and original data set is added, obtains
To the data set by expanding, while obtaining the 3D CNN model by pre-training.
Step 4: model being finely adjusted in the data by expanding, obtains final 3D MRI tumour voxel classification mould
Type.
(1) model framework is changed
By most latter two full articulamentum of pre-training 3D CNN model obtained by step 3, change convolutional layer into, wherein parameter with
Machine initialization, and standardized normal distribution is obeyed, 3D CNN model becomes 3D FCN model at this time.
(2) model is finely tuned
On the data set expanded by step 3.6, it is trained with the model that this step 4.1 obtains, training step
With reference to step 3, MRI tumour voxel classification model is finally obtained.
Preferred embodiment:
An optimal specific embodiment of the invention: the double convolution kernel 3D CNN network model frameworks of dual path, mould are established
Type is by 2 convolutional layers, 2 activation primitive layers, 2 pond layers, 2 standardization layers, 2 full articulamentums, 1 Dropout layers and 1
A objective function layer is successively constituted, and upper pathway uses big convolution kernel, and size is set as 21 × 21 × 11, and lower path uses rouleau
Product, size are set as 7 × 7 × 3.1 × 1 × 1 convolution kernel is added before second convolutional layer to reduce parameter and calculation amount.Instruction
The parameter of network is based on Xavier parameter initialization method when practicing network, carries out propagated forward, and class object function is based on
Softmax functional based method, backpropagation propagate undated parameter and are based on stochastic gradient descent method.Model loss, will not when changing
Original data set is added in false negative and false positive sample in classification results, obtains the data set by expansion.It is complete to change model again
Articulamentum is convolutional layer, it is made to become FCN, is then trained on the data set after expansion to the FCN model of pre-training, most
3D MRI tumour voxel classification model is obtained eventually.
A kind of MRI tumour voxel is identified using deep learning and machine vision technique in conclusion of the invention
Method establish the double convolution kernel 3D CNN model frameworks of dual path, upper pathway is using small first based on AlexNet model
Convolution kernel extracts the features such as texture, the size of focal area itself, and focal area and periphery are extracted using big convolution kernel in lower path
The features such as the correlation of tissue.1 × 1 × 1 convolution kernel is added before the convolutional layer of frame and carries out Feature Dimension Reduction, and is based on institute's structure
The model built is trained, by the false positive and the addition original data set progress sample expansion of false negative sample in result.Then by mould
Full articulamentum changes convolutional layer into type, i.e., CNN model is become FCN model, to pre-trained in the data expanded
The fine tuning of 3D FCN model, obtains final MRI tumour voxel classification model.The research reduces artificial detection medical imaging analysis
Cumbersome process, improve detection tumor image accuracy rate, to future medicine image automation and intelligent processing provide
Technical support.
In the description of this specification, reference term " one embodiment ", " some embodiments ", " illustrative examples ",
The description of " example ", " specific example " or " some examples " etc. means specific features described in conjunction with this embodiment or example, knot
Structure, material or feature are included at least one embodiment or example of the invention.In the present specification, to above-mentioned term
Schematic representation may not refer to the same embodiment or example.Moreover, specific features, structure, material or the spy of description
Point can be combined in any suitable manner in any one or more of the embodiments or examples.
Although an embodiment of the present invention has been shown and described, it will be understood by those skilled in the art that: not
A variety of change, modification, replacement and modification can be carried out to these embodiments in the case where being detached from the principle of the present invention and objective, this
The range of invention is defined by the claims and their equivalents.